基于Bayesian正则化BP神经网络的GPS高程转换  

GPS HEIGHT CONVERSION BASED ON BAYESIAN REGULARIZATION BP ARTIFICIAL NEURAL NETWORK

作　　者：张秋昭[1] 张书毕[1,2] 刘军[1] 王光辉[1] 王波[1] 

机构地区：[1]中国矿业大学环境与测绘学院,徐州221008 [2]江苏省资源环境信息工程重点实验室,徐州221008

出　　处：《大地测量与地球动力学》2009年第3期84-87,共4页Journal of Geodesy and Geodynamics

基　　金：国家自然科学基金(40401038)

摘　　要：针对标准BP神经网络算法泛化能力弱、易过度训练等问题,应用Bayesian正则化算法改进BP神经网络的泛化能力。通过对某矿区GPS联测水准点拟合计算,并与L-M算法、多项式曲面拟合等方法比较,Bayesian正则化的BP神经网络拟合精度更高、更稳定、泛化能力更强。In view of the defects of traditional BP artificial neural network, the Bayesian regularization method is used for improving the generalization ability of artificial neural network. Through an engineering example, the BP artificial neural network method with Bayesian regularization is compared with L-M method and polynominal curve fitting method respectively. It is found that the improved method has better fitting precision, stability and better generalization ability, so it has good practicability in GPS height conversion.

关 键 词：Bayesian正则化 BP神经网络 GPS高程转换 泛化能力 拟合 

分 类 号：P204[天文地球—测绘科学与技术]